Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.31563.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

A research paper applying Steiner's group task taxonomy to multi-agent LLM scaling, testing 13 open-weight models in teams up to 30 agents on disjunctive and compensatory tasks. Finds plurality voting fails to realize the potential of larger teams on disjunctive tasks, one peer revision matches 29, and averaging on Fermi estimation reduces error only ~6% due to shared model biases (~87% of squared error).

Quantifies when multi-agent LLM teams actually scale. The finding that a single peer revision matches 29 peers on accuracy is directly actionable for anyone designing agent workflows.

Cited by

No citations on record.